darthcrawl/mistral-7b-instruct-v0.3-artisan

TEXT GENERATIONConcurrency Cost:1Model Size:7BQuant:FP8Ctx Length:4kPublished:Apr 28, 2026License:apache-2.0Architecture:Transformer Open Weights Cold

darthcrawl/mistral-7b-instruct-v0.3-artisan is a 7 billion parameter LoRA fine-tune of Mistral-7B-Instruct-v0.3, developed by darthcrawl. This model is specifically optimized for direct, practical technical assistant responses and grounded character roleplay, maintaining broad general-purpose capabilities. It excels at providing concise technical answers without boilerplate and engaging in in-character dialogue with continuity and tone awareness. The model also includes an adult-adjacent content slice, making it more permissive than the base instruct model.

Loading preview...

darthcrawl/mistral-7b-instruct-v0.3-artisan Overview

This model is a 7 billion parameter LoRA fine-tune of mistralai/Mistral-7B-Instruct-v0.3, developed by darthcrawl. It was built on a curated dataset of hand-written technical Q&A and dialogue, balanced with general-purpose and conversational data to retain broad capabilities. The training utilized QLoRA with a 4-bit base and LoRA (r=16, alpha=32, dropout=0.05) targeting key projection layers, with a data mix of approximately 3% curated artisan, 14% character roleplay, and 83% general-purpose data over two epochs.

Key Capabilities

  • Concise Technical Answers: Provides direct and practical responses for topics like backend, systems, Go, databases, and networking, avoiding generic preambles.
  • Grounded Character Roleplay: Capable of engaging in in-character dialogue with strong continuity, restraint, and tone awareness, maintaining frame.
  • Adult-Adjacent Content: Includes a curated explicit-RP slice, making it more permissive than the stock instruct model.

Training Details & Limitations

The model was trained on a single H100 80GB GPU. It is primarily English-only and inherits the biases and knowledge cutoff of its base model. Quantized variants for MLX (Apple Silicon) are available, including 4-bit (vanilla and DWQ), 6-bit (near-lossless), and 8-bit (effectively FP16) versions. The model is licensed under Apache 2.0, inherited from the base model.

Finetunes
2 models